CFG Scale & Guidance in AI Image Generation (2026)

If your AI images look flat, burnt, or nothing like your prompt, the CFG scale is often the culprit. Classifier-Free Guidance — now usually just called guidance scale — is the single most powerful dial for controlling how literally a diffusion or flow-matching model follows your text. Get it right and your prompts snap into focus; get it wrong and you get plastic skin, crushed shadows, and oversaturated colors. This guide explains what CFG scale actually does, how the right values have shifted across model generations, and exactly how to tune it in 2026 workflows.

What Is CFG Scale?

CFG stands for Classifier-Free Guidance, a technique introduced in 2022 that lets a model compare two predictions at every denoising step:

• The conditional prediction — what the model thinks the image should look like given your prompt.
• The unconditional prediction — what it predicts with no prompt at all (or with only your negative prompt).

The guidance scale controls how far the final result is pushed away from the unconditional guess and toward the prompted one, using a simple extrapolation:

output = unconditional + scale × (conditional − unconditional)

A scale of 1 means “use the conditional prediction as-is.” A scale of 7 means “amplify the difference sevenfold.” That amplification is why high CFG produces exaggerated contrast, hard edges, and stiff compositions: you’re literally pushing the model’s latent math into extreme territory.

Why It’s Usually Called “Guidance” Now

Modern pipelines often rename the parameter. ComfyUI, Diffusers, and most APIs expose it as guidance_scale, while flow-matching models like Flux.1 and Stable Diffusion 3.5 use a guidance embedding — a distilled signal baked into the model during training rather than true classifier-free guidance. The effect feels similar to users, but the underlying mechanics differ, which is why the “correct” numbers are wildly different between model families.

How Guidance Changes Your Image

Think of CFG as a trade-off between prompt obedience and natural image statistics:

• Low CFG (1–4): Soft, painterly, sometimes dreamlike. High variety, but the model may ignore specific details or objects.
• Mid CFG (5–8): The sweet spot for most standard models. Good prompt adherence with believable lighting and texture.
• High CFG (10–20): Rigid, high-contrast, oversaturated. Useful for bold graphic styles, logos, or text, but prone to artifacts, halos, and anatomy errors.
• Very high CFG (20+): Almost always broken. Burned colors, repeated patterns, and shattered geometry.

Note that guidance also interacts with your negative prompt. Because the unconditional branch is where negatives live, CFG must be above 1 for negative prompts to have any effect at all. At CFG 1, your negative prompt is mathematically ignored.

Typical CFG Values by Model Family (2026)

There is no universal “correct” CFG. Always check the model card first, then adjust from there.

Model type	Typical guidance
SD 1.5 / 2.x checkpoints	7–8
SDXL base & most fine-tunes	4–7 (5 is a common base)
SD 3.5 Medium / Large	4–7
Flux.1 dev (guidance embedding)	3–4 (CFG set to 1)
Flux.1 schnell	0–1 (no guidance)
Turbo / LCM / Hyper / Lightning	1–2

Distilled and Turbo Models

Few-step models are trained to jump straight to a clean result in 1–8 steps. They rely on minimal or zero CFG, so pushing the scale to 7 will destroy the image. Keep guidance at 1–2 and let the sampler do the work.

Flow-Matching Models

Flux and SD 3.5 handle guidance differently. In ComfyUI, for example, Flux.1 dev runs with CFG at 1, while a separate FluxGuidance node (usually set between 2.5 and 4) controls prompt adherence. Treat this number like a lighter-touch version of CFG — 3.5 is the documented default, and going above 5 tends to flatten detail.

Closed Platforms

Midjourney, Ideogram, and hosted APIs increasingly hide the raw CFG value. Midjourney’s --stylize and --chaos parameters approximate the same trade-off, while OpenAI’s image models offer no direct control at all. When you can’t see the dial, prompt specificity and reference images become your guidance tools.

Practical Tips for Tuning CFG

These habits will save you hours of trial and error:

• Start at the model default, then move in 0.5 increments. CFG is sensitive; a jump from 6 to 8 is a bigger change than most people expect.
• Run XY plots. In Automatic1111 or ComfyUI, generate a grid sweeping CFG 3–9 with a fixed seed. You’ll see exactly where the image breaks down and where it loses prompt fidelity.
• Use CFG Rescale (0.6–0.8) for SDXL. This normalizes the guided prediction and tames the over-saturated, contrast-blasted look at higher guidance values without sacrificing prompt adherence.
• Don’t fix a bad prompt with higher CFG. If a concept is missing, describe it better — amplification won’t invent information the model doesn’t have.
• Match steps to guidance. Higher CFG usually needs more steps (25–40) to settle; low-CFG distilled models need only 4–8.
• Adjust one variable at a time. Changing CFG, sampler, and scheduler simultaneously makes results impossible to reason about.
• Test sampler compatibility. Some schedulers (DPM++ 2M Karras, Euler a) tolerate higher CFG better than others; ancestral samplers can amplify artifacts faster.

When to Push Guidance Higher

Raise CFG deliberately for graphic design work, typography, clean product mockups, and strongly symbolic compositions where sharpness and literal interpretation matter more than organic realism.

When to Pull It Lower

Lower CFG suits portraits, cinematic lighting, painterly illustration, and any scene where you want soft falloff, ambient depth, and creative interpretation. Many photographers’ best AI renders sit at CFG 3–5.

Common Mistakes to Avoid

• Assuming every model uses the same 7–8 range — a quick way to ruin Flux or Turbo output.
• Leaving CFG at 7 while running a 4-step distilled model.
• Blaming the model for artifacts that vanish at CFG 5.
• Ignoring negative prompts entirely, then wondering why CFG has no visible effect.
• Never revisiting old settings after a model update — defaults shift between versions.

Conclusion

CFG scale and guidance remain the most direct form of AI image control available to creators in 2026 — but the rules have fragmented. Standard SD 1.5 checkpoints still want 7–8, SDXL and SD 3.5 sit comfortably at 4–7, Flux.1 dev uses a low guidance embedding near 3.5, and distilled models expect 1–2. Treat the model card as your starting point, sweep values in half-point increments, and always evaluate guidance alongside your sampler, step count, and negative prompt. Master that balance and you stop fighting the model — you start directing it.